Fable's judgement

2026-07-03 · Source: Simon Willison's Weblog · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering · Depth: Intermediate, quick

Summary

On July 3rd, 2026, a key insight from an AIE Fireside Chat with Cat Wu and Thariq Shihipar from the Claude Code team highlighted the benefit of allowing large language models like Fable and Opus to exercise their own judgment. This approach, reinforced by Jesse Vincent, helps optimize token usage, especially before anticipated price increases. For instance, instead of explicitly dictating testing protocols, Fable can be instructed to decide when to write tests. A specific prompt, "For all coding tasks use your judgement to decide an appropriate lower power model and run that in a subagent," was used with Claude Code. This resulted in a memory file, `delegate-coding-to-subagents.md`, detailing how the main model delegates coding to subagents using Sonnet for substantive work or Haiku for trivial edits, while retaining judgment-heavy tasks. This strategy has proven effective in reducing Fable token consumption.

Key takeaway

For AI Engineers optimizing LLM operational costs, empower your top-tier models like Fable to use their judgment for task delegation. Instruct them to spawn subagents running lower-power models for routine coding, reserving the main model for complex review and synthesis. This approach significantly reduces token consumption, as demonstrated by the `delegate-coding-to-subagents` memory file, while maintaining quality oversight and efficiency.

Key insights

Empower LLMs to use their own judgment for task execution and model selection to optimize efficiency.

Principles

• Trust LLM judgment for task execution.
• Delegate routine tasks to lower-power models.
• Reserve top-tier models for complex judgment.

Method

Instruct the main LLM to use its judgment to select and run lower-power models in subagents for coding tasks, reviewing results in the main loop.

In practice

• Implement a memory file for delegation.
• Use Sonnet for substantive coding.
• Use Haiku for trivial code edits.

Topics

• Fable
• Claude Code
• LLM Cost Optimization
• Agentic AI
• Model Delegation
• Subagents

Best for: AI Engineer, Machine Learning Engineer, MLOps Engineer

Related on AIssential

• Fable 5 Raises the Bar for AI Ambition — Fable 5 shifts AI interaction from task delegation to responsibility delegation, demanding "task imagination" for complex, long-running projects.
• 🥇Top AI Papers of the Week — LLM internal states, external environments, and architectural choices profoundly impact performance, safety, and cost.
• Claude Fable 5 is here! — Claude Fable 5 offers unparalleled performance for complex, long-horizon tasks at a premium cost.
• Mythos is here, it’s time to start tokenmaxxing — Maximizing subsidized AI model usage requires strategic rate limit management and multi-agent orchestration for complex tasks.
• Turning Andrej Karpathy’s LLM Coding Thoughts into Claude.md — LLMs are transforming coding workflows, shifting developers from direct coding to guiding and reviewing AI agents.
• Why Fable 5 (Mythos 5) doesn't need Agents at all — The Fable 5 architecture achieves complex cognition by externally reconstructing state and behavior for a frozen LLM via a recursive harness and artifacts, not internal agents.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Simon Willison's Weblog.